A better architecture might be that the language model can determine possible methods to solve mathematical problems？Has anyone thought about this direction?

A better architecture might be that the language model can determine possible methods to solve mathematical problems, then perform specific operations to transform the solution of the mathematical problem into a series of thoughts, fill the thought chain, and add generality in the transformation part to make the thinking more divergent, and then the part of filling the thought chain needs to be more accurate

It's a bit like the architecture of chatgpt+mathematica, but the problem with the chatgpt+mathematica architecture is that mathematica is too biased towards hard coding and often reports errors, and chatgpt does not specifically train to split the input of a mathematical problem into step-by-step solutions